|Publication number||US7702931 B2|
|Application number||US 11/475,418|
|Publication date||Apr 20, 2010|
|Filing date||Jun 27, 2006|
|Priority date||Jun 27, 2006|
|Also published as||US20070300083|
|Publication number||11475418, 475418, US 7702931 B2, US 7702931B2, US-B2-7702931, US7702931 B2, US7702931B2|
|Inventors||Alan L. Goodrum, Roger E. Tipley|
|Original Assignee||Hewlett-Packard Development Company, L.P.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (35), Non-Patent Citations (13), Referenced by (35), Classifications (8), Legal Events (3)|
|External Links: USPTO, USPTO Assignment, Espacenet|
Computers are ubiquitous in society. For example, computers are present in everything from user-oriented desktop computer systems to complex networks of computers that facilitate credit card transactions. These complex networks represent a trend toward consolidating computers to implement high-density computing configurations, which are sometimes referred to as “data centers.” In fact, it is not uncommon for these data centers to include tens of thousands of servers or more. To support of these data centers, information technology (IT) professionals have had to shoulder new burdens that were previously not of concern to IT professionals: power consumption and temperature maintenance.
Previously, data center facilities managers were primarily responsible for providing the specified power to the computers within the data center and were also responsible for maintaining the ambient air conditions to match the specified operating conditions of the computers within the data center. Typically, power and cooling requirements of the data center were estimated based on the “name plate” ratings—an estimate of the power and cooling requirements provided by the computer manufacturer. Historically, when computer server power levels were low, this approach proved practical. More recently, however, IT professionals have begun implementing servers as “blade servers” where each chassis is filled with multiple server modules, or “blades.” Increasing the density of the servers in this manner may result in cooling costs for some data centers (e.g., 30,000 ft2) running into the tens of millions of dollars per year and also may result in a higher incidence of undesirable service outages caused by cooling failures.
Some facilities and IT professionals have begun to “de-rate,” or reduce the name plate power and cooling requirements by a fixed amount to increase the density of servers within a data center. De-rating, however, may still undesirably mis-predict actual power consumption. Other attempts to increase server density include estimating the actual power consumed with an anticipated application running on the server to provide a more accurate estimate of the actual power requirements of the server. However, an unexpected change in computing demand, for example as a result of a work load shift between servers, may increase power demand and trip a circuit breaker or cause localized over-heating.
For a detailed description of exemplary embodiments of the invention, reference will now be made to the accompanying drawings in which:
Certain terms are used throughout the following description and claims to refer to particular system components. As one skilled in the art will appreciate, computer companies may refer to a component by different names. This document does not intend to distinguish between components that differ in name but not function. In the following discussion and in the claims, the terms “including” and “comprising” are used in an open-ended fashion, and thus should be interpreted to mean “including, but not limited to . . . . ” Also, the term “couple” or “couples” is intended to mean either an indirect or direct electrical connection. Thus, if a first device couples to a second device, that connection may be through a direct electrical connection, or through an indirect connection via other devices or connections.
The term “blade server” is intended to refer to a computer system with one of its dimensions, usually the width, substantially smaller than the rest. This is usually accomplished by integrating a majority of the server's components (including processor(s), memory, network cards, etc.) onto the motherboard, allowing multiple blade servers to be rack mounted within a common housing enclosure.
The term “management device” is intended to refer to any device that permits remote control functionality such as remote system configuration, remote power control, remote system reset, or remote operating system (OS) console of a host computer system.
The term “U” is intended to refer to a standard unit of measure for the vertical space that servers occupy within a server enclosure. Server enclosures such as racks and cabinet spaces as well as the equipment that fit into them are usually measured in U. 1 U is equivalent to about 1.75 inches. Thus, a rack designated as 20 U, is a server enclosure that has 20 rack spaces for equipment and housing enclosures and has 35 (20×1.75) inches of vertical usable space.
The term “power-regulation state,” sometimes called “p-states,” is intended to refer to varying levels of power consumption by the CPU, where each level or p-state indicates a different level of CPU functionality.
The term “data center” is intended to refer to a group of computers that perform a computing function. The computers from these data centers are often, but not always, housed in a common building and may include thousands of computers.
The term “clock modulation” is intended to refer to stopping the clock signal provided to some or all of the central processing unit (CPU) and/or other portions of a computer that may share this clock. Clock modulation is often achieved by asserting a STPCLK signal pin of the CPU or its support chipset.
The following discussion is directed to various embodiments of the invention. Although one or more of these embodiments may be preferred, the embodiments disclosed should not be interpreted, or otherwise used, as limiting the scope of the disclosure, including the claims. In addition, one skilled in the art will understand that the following description has broad application, and the discussion of any embodiment is meant only to be exemplary of that embodiment, and not intended to intimate that the scope of the disclosure, including the claims, is limited to that embodiment.
Allocating Power within a Data Center
The current theoretical limit of allowable power density is believed to be about 1000 watts per square foot (W/ft2) including the vacant aisle space between various server enclosures. This theoretical limit of 1000 W/ft2 assumes that water chilled cabinets are used to house the servers, although some embodiments may implement cooling techniques that yield a higher theoretical limit. Thus, even if an elaborate cooling system is implemented, if a server rack and the unoccupied floor space around it take up 18 square feet (ft2) of floor space, for example, then that rack is theoretically limited to 18,000 watts (W) of power. A typical 1 U server may consume on the order of 620 W/server, and therefore, the maximum number of servers that may be implemented in this example is approximately 29, whereas a typical rack may be capable of physically accommodating forty-two 1 U servers. Thus, despite having the floor space within data center 100 to add more servers to the rack, and the need to increase computing power by adding more servers to the rack, businesses may be limited from filling the racks to their full capacity.
The embodiments of the present invention may allow for this additional server capacity by budgeting the power allocated to “hypothetical” levels within the data center. These hypothetical levels represent user-defined entities that have different power consumption requirements. By actively managing the power consumed at each of these hypothetical levels, the total power consumed by the data center may be kept within a desired budget, thereby allowing more servers to be added to the data center.
One key feature of zone manager 107 is the ability to adaptively monitor and curtail the power consumption in a zone within the predetermined amount allocated to it by data center manager 106. For example, if zone manager 107 monitors physical conditions in zone A and determines that HVAC unit 104 cannot cool zone A—e.g., because the physical placement of racks 105 within zone A does not facilitate good airflow—then the zone manager 107 located in zone A may actively reduce the power budget assignment of racks 105 within zone A and report this information to data center manager 106.
In a recursive manner, racks 105 also may include a rack manager 108, which may be defined as another hierarchical level of abstraction within software such as HP SIM. Although rack manager 108 is depicted in
During operation, rack manager 108 may receive a power budget for its rack from zone manager 107. In turn, rack manager 108 may divide up this power budget among server enclosures 112. Likewise, each enclosure 112 may include an enclosure manager 111 that is capable of dividing the power budget for enclosure 112 among the plurality of servers 110 within enclosure 112. While enclosure manager 111 is shown housed in the same enclosure 112 as servers 110, other embodiments include housing enclosure manager 111 in other locations. In fact, in some embodiments, enclosure manager 111 may be implemented in software such as HP SIM.
Further still, each server 110 may include a management processor 114 that is capable of limiting the amount of power consumed by server 110 to be less than or equal to the dynamically assigned budget from enclosure manager 111. Thus, if zone manager 107 reduces the power budget of zone A, then one or more rack managers 108 may further reduce the power budget on one or more enclosure managers 111, which in turn may reduce the power budget for one or more management processors 114. Note that this adjustment by zone manager 107, rack managers 108, and/or enclosure managers 111 may occur independently and therefore a reduction in the power budget of zone A is not required prior to a reduction of the power budget of a particular rack within zone A.
In the end, management processor 114 may react to this reduction in power budget by limiting the amount of power available to individual components within server 110. Actions that may be taken by server 110 to curtail its power consumption within the budgeted amount include: slowing down server 110, off-loading work from server 110 and then shutting it down, not allowing server 110 to be powered up, or off-loading work such that the server spends more time in the idle state. For example, management processor 114 may move work load to a server that consumes less power or is in another zone. Equations 1-4 represent the mathematical expressions for the power budgets of the various levels of data center 100.
Equation 1 illustrates that the sum of the power budgets of each zone is less than or equal to power budget assigned to the entire data center. Equation 2 demonstrates that the sum of the power budgets for each rack (e.g., Rack A through Rack Z) within zone A is less than or equal to the total assigned power budget for zone A. Equation 3 illustrates that the sum of the power budgets for each server enclosure within the rack is less than or equal to the total assigned power budget for a rack. Equation 4 demonstrates that the sum of the power budgets of each of the individual servers is less than or equal to the total assigned power budget for the server enclosure. Although four levels of power budget allocation and control are described in Equations 1-4, in practice, there may be multiple levels of power budget allocation and control (i.e., at least one at each hierarchical level).
Automatic Adjustment of Power Budgets Within a Data Center
As was alluded to above, each server 110 may track (e.g., using management processor 114) its power budget assignment. In addition to tracking its power consumption, however, server 110 also may negotiate to adjust its power budget as its present needs change. This negotiation may occur between various components within data center 100, such as between data center manager 106 and zone managers 107. For example, data center manager 106 may direct zone manager 107 in zone B to use any excess power reported by zone manager 107 in zone A.
Non volatile storage 311 is capable of storing executable code and data. The contents of non-volatile storage 311 may be changed from time to time by reprogramming either a portion or the entire non-volatile storage 311.
Bridge logic device 312 may be referred to as a “North bridge.” In some embodiments, bridge 312 couples to a memory 314 by a memory bus (M-BUS). In other embodiments, however, CPU 310 includes an integrated memory controller, and memory 314 connects directly to CPU 310.
Bridge 312 also couples to PCI-Express® slots 318A-B using the PCI-Express® bus standard as disclosed in “PCI-Express Base Specification 1.0a,” available from the PCI Special Interest Group (PCI-SIG) and incorporated herein by reference.
As noted above, server 302 may be implemented as a blade-type server that is part of a larger data center, such as data center 100. Regardless of the actual implementation of server 302, a management processor 330 may be included in server 302. Management processor 330 couples to the various portions of server 302 as well as coupling to power managers for server enclosures 108 as shown. In some embodiments, management processor 330 couples directly to North Bridge 312 via PCI or PCI-Express bus, and in other embodiments, management processor 330 couples to North Bridge 312 via a combination of a South Bridge 320 and a PCI-Express bus. Commercial implementations of management processor 330 include Hewleft-Packard's Integrated Lights Out (iLO) processor.
During operation, management processor 330 tracks the amount of power assigned to it by server enclosure 108 (shown if
CPU utilization refers to a measure of how much of a CPU's computing capacity is being used and indicates the overall activity level of a server. Fundamentally, CPU utilization may be thought of as the percentage of time that the CPU spends in a non-idle, active state. For example, a CPU that is 100% utilized is executing its maximum possible workload.
The CPU's power-regulation state, sometimes termed “p-states,” refers to varying “states” or levels of power consumption by the CPU, where each level or p-state indicates a different level of CPU functionality. Since the CPU is made of numerous transistors that switch on and off to perform desired functions, and each time these transistors switch power is consumed, the faster the CPU operates the more power it will consume. Accordingly, the different p-states may be accomplished by adjusting the operating frequency of the CPU. (Note that altering the clock frequencies to achieve different p-states should not be confused with “clock modulation,” which is described in more detail in subsequent sections of this disclosure.) In addition, to being proportional to the operating frequency the amount of power consumed by the CPU is also proportional to the square of the CPU's operating voltage. In other words, the lower the operating voltage, the lower the amount of power consumption. Therefore, different p-states may have different operating frequencies and different operating voltages.
For example, the P0 p-state is recognized as the highest level of performance (and highest possible level of power consumption) since the CPU runs at full operating frequency and full operating voltage in the P0 p-state. Thus, p-states lower than P0, which may be referred to as P1, P2, etc., will include either a lower operating frequency or a lower operating voltage or both. The precise value of the operating voltage and operating frequency for each p-state is determined by the CPU vendor to optimize CPU utilization while minimizing overall power consumption. Furthermore, each p-state has a target utilization, e.g., P0 may have a 100% utilization target while P1 may have a 70% utilization target. In accordance with some embodiments, each p-state may be associated with an upper threshold above this utilization and a lower threshold stored in non-volatile storage 311 and these thresholds may be used to adjust the p-state of the CPU as described in more detail below in the context of
In any case, management devices, such as management processor 330 or enclosure manager 111, may use p-states to manage the power requirements of server 302. For example, assume that server 302 is in a p-state other than P0 and that CPU utilization is at or near 100%. Further, assume that according to embodiments of the present invention, server 302 is prevented from raising its p-state (e.g., from a lower p-state to the P0 p-state) because of an assignment from an entity further up in the hypothetical levels of data center 100, such as a power state assignment from a server enclosure manager 111. In this scenario where the CPU is 100% utilized and its p-state can be raised, server 302 could benefit from an increase in its power budget, and therefore may request that enclosure manager 111 increase its power budget. In a similar fashion, if CPU utilization drops so that the CPU would be less than 100% utilized in the next lower p-state, then server 302 may request that enclosure manager 111 reduce the power budget assigned to server 302, which in turn may request that rack manager 108 to reassign a power allocation.
This power budgeting process between the various levels of data center 100 continues in a recursive fashion. That is, as enclosure manager 111 receives requests for power from the various servers that it manages, if the total power requested by these servers exceeds the power budget set by rack manager 108 (i.e., the next level up in the hierarchy of data center 100), then enclosure manager 111 will request additional power from rack manager 108.
Since this negotiation process for more or less power budget varies with CPU utilization, and since CPU utilization is unknown prior to booting up server 302, a baseline power budget may be helpful. Accordingly, prior to powering on server 302, management processor 330 may be programmed with the name plate power requirements, and name plate power requirements may be used to power on and boot up server 302. After powering on server 302, however, the amount of power for operation may decrease, and therefore management processor 330 may reassess the power requirements of server 302 and adjust accordingly.
With the maximum power rating known, management processor 330 then asks permission from the enclosure manager 111 (or other managers higher up the hierarchical chain) to startup server 302 per block 351. The enclosure manager 111 then determines whether allowing server 302 to startup will cause the enclosure manager 111 to exceed its power budget in block 352. If allowing server 302 to startup will cause enclosure manager 111 to exceed its power budget, then algorithm 340 may loop back to block 350 and not allow server 302 to startup until either the power budget for enclosure manager 111 changes or the maximum power requirement for server 302 changes. On the other hand, if allowing server 302 to startup will not cause enclosure manager 111 to exceed its power budget, then server 302 is then initialized and a power-on self test (POST) may be performed in block 355. Note that during execution of blocks 350-355, management processor 330 is operational despite the fact that server 302 may not be operational.
During boot up, management processor 330 may renegotiate a power budget that is less than the name plate rating stored in memory, per block 360. In some embodiments, this negotiation may take place between management processor 360 and other management devices, such as enclosure manager 111 and zone manager 107. This negotiation process may include a scheme among servers that prioritizes the order that servers give power back to the overall power budget and also prioritizes the order that servers take power from the overall power budget. For example, some servers may be executing critical applications and therefore they may be given higher priority than servers executing non-critical applications. Further still, the negotiation process may include staging servers such that additional servers do not power on and begin negotiating a power budget until servers with a higher priority have completed booting up and have reached a stable power budget. This negotiation process also may be based on previous history (which may be stored in management processor 330) of servers that have given up and taken back power from the power budget in the past.
Once the negotiation is compete, in block 365, management processor 330 accounts for the power either provided to or taken from the total power budget. Management processor 330 may then report this accounting to enclosure manager 111, or other components higher up in hierarchy. Although the boot process may be complete (in block 370) and control may have been passed off to the operating system running on server 302, management processor 330 may dynamically reassess the power requirements of server 302 by looping through blocks 360 and 365 during server operation as is illustrated.
During server operation, management processor 330 may determine (in block 375) if the current CPU utilization is above or below the upper and lower utilization thresholds for the current p-state. This may occur, for example, by checking the contents of non-volatile storage 311. If the current CPU utilization is above the upper utilization threshold for the current p-state then management processor 330 may determine if there is a higher p-state available and may adjust the p-state of CPU 310 accordingly per block 380. At this point, management processor 330 may repeat algorithm 340, including block 365, to account for any power increases within the total power budget.
If, however, the current CPU utilization is below the utilization for a p-state, then in block 385, management processor 330 selects a p-state (which may be based on the thresholds stored in non-volatile storage 311) whose upper and lower utilization thresholds match the current server utilization. Effectively, if a lower p-state is chosen as a result of block 385 then CPU 310 is relinquishing at least some of the power allocated to it by management processor 330. Thus, according to block 360, management processor 330 may negotiate with other management devices (e.g., enclosure managers 111 or zone managers 107) to distribute the amount of power relinquished by CPU 310 among the other devices under the control of that particular management device. For example, if management processor 330 is negotiating with enclosure manager 111 then other servers within the same enclosure may receive the relinquished power, whereas if management processor 330 is negotiating with zone manager 107, then other servers within the same zone may receive the relinquished power.
In order to prevent the server that relinquished this power from having to renegotiate this power back if needed, some embodiments restrict management processor 330 to only negotiate with specific entities in the hypothetical chain, such as with other servers in the same enclosure manager 111. This may be particularly useful to prevent thrashing—i.e., where the server relinquishing power goes back and forth between periods of high activity (i.e., requires more power) to periods of low activity (i.e., requires less power).
In addition to reducing the power consumption of the server by reducing the power consumption of the CPU, other system components that rely on the number of requests from CPU (e.g., memory or disk drives) also may have their power reduced as a result of reducing the power consumption of the CPU. For example, during operation, the CPU makes disk access requests of the hard drive. Therefore, as the operating frequency of the CPU decreases because of lowering p-states, the number of disk access requests also may decrease, and hence, the hard disk may consume less power.
Power Budgets Based on Estimated Power Consumption
As mentioned previously, the name plate power may be used as a baseline power budget upon start up. Note that the name plate power is usually a general estimate for each model of server regardless of the server's configuration. In some embodiments, however, an estimated power consumption of server 302 may be used as a baseline power budget instead of the name plate rating. Such a feature may be useful in that servers may be configured differently (e.g., different number of hard drives) and a power estimate of the actual server being managed may represent a more accurate power budget to begin with.
In order to fully regulate the power consumption of the server, an additional factor that may be used to control CPU power consumption is the CPU's “clock modulation,” which is a power control factor that is separate and apart from CPU p-state and CPU utilization. Clock modulation is where the clock frequency provided to some or all of the CPU is stopped for a period of time, which substantially reduces power consumption to the portions of the CPU that have their clock stopped. As is evident to one of ordinary skill in the art, the term “STPCLK” is an industry standard term for the CPU clock modulation control signal.
In some embodiments, the CPU may include a crystal oscillator that provides a base frequency to the CPU, and this base frequency may be increased (possibly by an internal phase locked loop) and then provided to other blocks within the CPU. In these embodiments, clock modulation may include stopping the clock frequency at the crystal oscillator, stopping the clock frequency at the phase locked loop, or both. Regardless of the actual internal clock distribution within the CPU, the CPU itself may include a STPCLK connection such that when STPCLK is asserted some or all of the internal CPU clock is stopped. Note that the functionality of STPCLK may be either active high or active low, and therefore, “asserting” STPCLK may include coupling a low signal to the STPCLK connection in some embodiments and in other embodiments it may include coupling a high signal to the STPCLK connection. By controlling the duty cycle, or percentage of time that a signal coupled to the STPCLK connection is asserted, power regulation may be achieved through clock modulation.
This clock modulation, in addition to other server settings may be used to regulate the power budget, where these settings collectively are referred to herein as Server Power Performance States or “SPP-states.” That is, SPP-states represent a combination of settings within a server to effectuate a predetermined amount of power consumption. The constituent SPP-state settings include, but are not limited to, CPU p-states, CPU clock modulation or STPCLK settings, as well as various configurations for the server's subsystem components (e.g., the speed at which a hard disk drive rotates).
The SPP-state settings for a particular server may be determined prior to deployment and stored in non-volatile storage 311 so that the appropriate STPCLK, p-state, and subsystem settings can be made that limit power consumption within the power budget set by the enclosure manager. For example, server 302 may be outfitted with various available options such as different hard drive types and sizes, memory types and sizes, network cards, and power supplies. In order to measure power consumption for each of these unique hardware combinations, maximum computational work loads may be run on the server. These maximum workload tests are deliberately chosen to force server 302 to consume as much power as possible. For example, specific software code that activates some of the more power hungry portions of CPU 310, like the floating point unit or arithmetic logic unit, may be run on server 302. During this maximum workload test, the SPP-state settings that result in different power consumption levels also may be determined. These SPP-state settings that effectuate different power levels are then stored and this information is made available to hardware or software on server 302.
Referring again to
Power supply 402 further couples to a power measurement circuit 404. Power measurement circuit 404 measures the power consumption of server 400, which may include CPUs 406 and additional system components 408 (e.g., memory or hard drives). A comparison circuit 410 couples to power measurement circuit 404 as well as coupling to a register 411. Register 411 may include a power budget value from a management processor 412 (indicated by the dashed line in
The power measurement from measurement circuit 404 is fed to comparison circuit 410 and therefore may be referred to as a closed-loop approach. In other embodiments that may be referred to as a more open-loop approach, comparison circuit 410 may receive a power estimate based on predetermined lab characterizations instead of real time measurements. In yet other embodiments, a hybrid approach may be used where comparison circuit 410 uses the lab characterization at first to start server 400, and then builds a lookup table 414 with the power measurement values from measurement circuitry 404 as server 400 operates and then uses this instead of the lab characterization data.
Regardless of the source of the power consumption, comparison circuit 410 may compare this power consumption to the power budget stored in register 411 and couple the result of this comparison to support logic 413. In turn, support logic 413 may control a STPCLK signal that is fed to the one or more CPUs 406 to stop at least some of the clock signals within CPUs 406. For example, if the power consumption of server 400 is much greater than the power budget value stored in register 411, then STPCLK may have a relatively high duty cycle so that the amount of power consumed by CPUs 406 is reduced. Likewise, if the power consumption is less than or equal to the value stored in register 411, then STPCLK may have a relatively low duty cycle (e.g., 0% stopped) so that the amount of power consumed by CPUs 406 is not restrained by STPCLK. In addition to comparison circuit 410, management processor 412 may adjust other portions of the SPP-state, such as the p-states of CPUs 406 or the parameters related to the additional system components 408 (e.g., hard drive rotation speed).
Determining Actual Power Dissipation for SPP-states
While the embodiments of the present invention shown in
As shown in the legend, each curve in
In block 502, the server is powered on and begins the power on self test (POST) procedure. In general, POST includes code that is executable by the server and is stored in a storage location, such as non-volatile storage 311. POST may include initializing the server as well as the server's peripherals per block 504. Once the initialization is complete, the server then performs a worst case synthetic work load test in block 506. During this synthetic work load test, non-volatile storage 311 may execute code that reads the actual power consumption as measured by measurement circuitry 404 (shown in
The values associated with each of these SPP-states may be entered into lookup table 414 (shown in
Prior to constructing lookup table 414, the overall accuracy of algorithm 500 may be set in block 508. If the accuracy of algorithm 500 is set high, then the work load test of block 506 may be repeated for each defined SPP-state per block 510. Once each SPP-state is defined, the server proceeds to booting the operating system in block 516.
On the other hand, if the desired accuracy of algorithm 500 is set low, then algorithm 500 generates fewer points for lookup table 414 and management processor 412 interpolates between these points as illustrated in block 512. For example, in situations where it is unnecessary to calculate the SPP-state with high accuracy, then the code in non-volatile storage 311 may measure the worst case power at SPP0 and also measure the power consumed by the CPU in the P0 p-state while the CPU is idle, i.e., the minimum power consumption, and then management processor 412 may interpolate between these two by mathematical interpolation. With this interpolation complete, in block 514, algorithm 500 constructs lookup table 414. With lookup table 414 constructed, the server proceeds to booting the operating system as illustrated in block 516. In some embodiments, these SPP-states may be determined each time the server is initialized and prior to booting the operating system.
Adjustment of Power Levels During Operation
As discussed above, each server in the data center may be required to limit its power consumption to within a power budget value assigned by a management device higher up in the hierarchy. In the closed-loop embodiments of the present invention, this may be accomplished by choosing values for the constituent settings of the SPP-states. In these closed-loop embodiments, each SPP-state may be associated with constituent settings such as p-state, STPCLK, and the settings of additional system components 408. As discussed above in the context of
Accordingly, algorithm 600 illustrated in
Referring now to
While the server is operating at this initial power level, in block 604, it is determined whether power regulation action was required, during a closed loop operation, to keep server 400 from exceeding the power budget established by management processor 412. For example, the user's actual instruction sequence may consume more power than the preselected instruction sequence of the initial SPP-state. If no regulation was required to keep server 400 from exceeding the power budget, then block 604 repeats to continually monitor whether regulation action has been required. On the other hand, if power regulation was required to keep server 400 from exceeding the power budget, then this indicates that the server's operating conditions have changed such that the closed-loop settings may not represent the most optimum value, and a flag may be set per block 606. This flag may be implemented in hardware as a bit in a register of management processor 412 or alternatively this flag may be implemented in software.
Management processor 412 periodically checks this flag, and if set, increases the power level associated with the current SPP-state per block 608. This may occur through increasing the constituent settings of the SPP-state, such as the p-state. Along with increasing the power level associated with the SPP-state in block 608, management processor 412 resets the maximum power measurement of block 602.
Algorithm 600 then determines whether the CPU (such as CPUs 406 shown in
The above discussion is meant to be illustrative of the principles and various embodiments of the present invention. Numerous variations and modifications will become apparent to those skilled in the art once the above disclosure is fully appreciated. For example, although servers are mentioned throughout this disclosure other devices within a data center, such as switches and routers, also fall within the scope of this disclosure. In addition, while power thresholds may have been discussed in this disclosure, this disclosure also extends to energy thresholds that measure power consumption over time. In this manner, the disclosed embodiments may be regulated according to the amount of heat that they produce. It is intended that the following claims be interpreted to embrace all such variations and modifications.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5446852||Aug 1, 1994||Aug 29, 1995||Deutsche Itt Industries Gmbh||Method of reducing noise signals caused by fluctuations in power requirements of digital signal processor having recurring wait routines|
|US5483656 *||Jan 14, 1993||Jan 9, 1996||Apple Computer, Inc.||System for managing power consumption of devices coupled to a common bus|
|US5621343||Sep 12, 1995||Apr 15, 1997||Mitsubishi Denki Kabushiki Kaisha||Demodulator circuit which demodulates pulse width modulated signals used in a semiconductor integrated circuit|
|US6442700||Aug 10, 1999||Aug 27, 2002||Intel Corporation||Thermal control within systems having multiple CPU performance states|
|US6711691||May 8, 2000||Mar 23, 2004||Apple Computer, Inc.||Power management for computer systems|
|US6775787||Jan 2, 2002||Aug 10, 2004||Intel Corporation||Instruction scheduling based on power estimation|
|US6775997||Oct 3, 2002||Aug 17, 2004||Hewlett-Packard Development Company, L.P.||Cooling of data centers|
|US6795928||Mar 18, 2002||Sep 21, 2004||International Business Machines Corporation||Method for managing power consumption of multiple computer servers|
|US6802014 *||Oct 26, 2000||Oct 5, 2004||Apple Computer, Inc.||Method and apparatus for managing power in computer systems|
|US6904534||Sep 29, 2001||Jun 7, 2005||Hewlett-Packard Development Company, L.P.||Progressive CPU sleep state duty cycle to limit peak power of multiple computers on shared power distribution unit|
|US7032119||May 18, 2001||Apr 18, 2006||Amphus, Inc.||Dynamic power and workload management for multi-server system|
|US7043647 *||Sep 28, 2001||May 9, 2006||Hewlett-Packard Development Company, L.P.||Intelligent power management for a rack of servers|
|US7058826||Oct 4, 2004||Jun 6, 2006||Amphus, Inc.||System, architecture, and method for logical server and other network devices in a dynamically configurable multi-server network environment|
|US7111177||Oct 25, 2000||Sep 19, 2006||Texas Instruments Incorporated||System and method for executing tasks according to a selected scenario in response to probabilistic power consumption information of each scenario|
|US7155623 *||Dec 3, 2003||Dec 26, 2006||International Business Machines Corporation||Method and system for power management including local bounding of device group power consumption|
|US7203849 *||Dec 12, 2003||Apr 10, 2007||Hewlett-Packard Development Company, L.P.||Method and system for distributing power to networked devices|
|US7272517||Apr 25, 2006||Sep 18, 2007||International Business Machines Corporation||Method and system for providing performance estimations for a specified power budget|
|US7400062 *||Oct 31, 2005||Jul 15, 2008||Microsemi Corp. - Analog Mixed Signal Group Ltd.||Rack level power management|
|US7421601||Feb 17, 2006||Sep 2, 2008||International Business Machines Corporation||Method and system for controlling power in a chip through a power-performance monitor and control unit|
|US7467306||Mar 8, 2005||Dec 16, 2008||Hewlett-Packard Development Company, L.P.||Methods and systems for allocating power to an electronic device|
|US7509506 *||Jun 9, 2005||Mar 24, 2009||International Business Machines Corporation||Hierarchical system and method for managing power usage among server data processing systems|
|US20020188877||Jun 7, 2001||Dec 12, 2002||Buch Deep K.||System and method for reducing power consumption in multiprocessor system|
|US20030009705||Feb 21, 2002||Jan 9, 2003||Michael Thelander||Monitoring and synchronization of power use of computers in a network|
|US20030055969||Sep 17, 2001||Mar 20, 2003||International Business Machines Corporation||System and method for performing power management on a distributed system|
|US20030105984 *||Dec 3, 2001||Jun 5, 2003||Dell Products L.P.||System and method for autonomous power sequencing|
|US20050268189||May 28, 2004||Dec 1, 2005||Hewlett-Packard Development Company, L.P.||Device testing using multiple test kernels|
|US20060168461||Jan 6, 2006||Jul 27, 2006||Allen Daniel J||Power management of components having clock processing circuits|
|US20060230299||Apr 11, 2005||Oct 12, 2006||Zaretsky Lee B||System and method for power usage level management of blades installed within blade servers|
|US20060282685 *||Jun 9, 2005||Dec 14, 2006||International Business Machines Corporation||Distributed system and method for managing power usage among server data processing systems|
|US20070067657 *||Sep 22, 2005||Mar 22, 2007||Parthasarathy Ranganathan||Power consumption management among compute nodes|
|US20070283176||Jul 3, 2007||Dec 6, 2007||Advanced Micro Devices, Inc.||Method and apparatus for improving responsiveness of a power management system in a computing device|
|US20080242338||Mar 26, 2008||Oct 2, 2008||Takeshi Yamamoto||Radio communication apparatus and transmission power control method|
|US20090049317 *||Aug 17, 2007||Feb 19, 2009||Alan Gara||Managing Power in a Parallel Computer|
|US20090113221 *||Oct 29, 2007||Apr 30, 2009||Microsoft Corporation||Collaborative power sharing between computing devices|
|US20090125293 *||Nov 13, 2007||May 14, 2009||Lefurgy Charles R||Method and System for Real-Time Prediction of Power Usage for a Change to Another Performance State|
|1||Advanced Micro Devices, Inc., "Thermal Management & Server Density: Critical Issues for Today's Data Center," White Paper, Copyright 2004 Rackable Systems, 8 pp.|
|2||Advanced Micro Devices, Inc., Acer Ferrari 4005 WLMi AMD Turion64 ML-37 Laptop Review, Copyright 2006 PCstats.com, 3 pp.|
|3||Chris Nuttall, "Intel set to unveil power saving processor," Aug. 14, 2005, 3 pp.|
|4||Enea Embedded Technology, "OSE Real-time Operating System to Support Freescale i.MX21 Applications Processor," Copyright 2006 Enea, 4 pp.|
|5||Helena Purgatorio, "Frequency switching improves power management in Power Architecture chips," Sep. 27, 2004, Updated Oct. 15, 2004, 3 pp.|
|6||Intel Corporation, "Addressing Power and Thermal Challenges in the Datacenter," White Paper, Jan. 2005, 12 pp.|
|7||Justin Moore et al., "Making Scheduling "Cool": Temperature-Aware Workload Placement in Data Centers," USENIX 2005 Annual Technical Conference, General Track, 16 pp.|
|8||Mel Tsai, "Inside DSP on Low Power: Designing Low-Power Signal Processing Systems," Jun. 1, 2004, 10 pp.|
|9||Microsoft Corporation, "The Laptop Computer May Be Unable to Enter the C3 Processor Power-Saving State," Article ID 297045, Revision 2.2, Jan. 27, 2006, 3 pp.|
|10||Microsoft Corporation, "The processor never switches to the C3 power-saving state when you remove a USB floppy drive from a Toshiba Tecra 9000 portable computer," Article ID 315754, Revision 1.1, Sep. 27, 2005, 2 pp.|
|11||Ratnesh K. Sharma et al., "Balance of Power: Dynamic Thermal Management for Internet Data Centers," Feb. 18, 2003, 14 pp.|
|12||Samer Al-Khairi, "Power Management Features On The Embedded Ultra Low-Power Intel486 Processor," Copyright 2005 Intel Corporation, 5 pp.|
|13||Win Guides Network, "Increase the USB Polling Interval," last modified Oct. 30, 2002, 2 pp.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7984311 *||Aug 8, 2008||Jul 19, 2011||Dell Products L.P.||Demand based power allocation|
|US8032768 *||Oct 4, 2011||Dell Products, Lp||System and method for smoothing power reclamation of blade servers|
|US8046600 *||Oct 29, 2007||Oct 25, 2011||Microsoft Corporation||Collaborative power sharing between computing devices|
|US8276012||Jun 30, 2009||Sep 25, 2012||International Business Machines Corporation||Priority-based power capping in data processing systems|
|US8375228 *||Aug 6, 2008||Feb 12, 2013||International Business Machines Corporation||Multiple-node system power utilization management|
|US8381000||Jun 15, 2011||Feb 19, 2013||Dell Products L.P.||Demand based power allocation|
|US8478451 *||Dec 14, 2009||Jul 2, 2013||Intel Corporation||Method and apparatus for dynamically allocating power in a data center|
|US8495393 *||Jul 22, 2009||Jul 23, 2013||Electronics And Telecommunications Research Institute||Remote power management system and method for cluster system|
|US8589556||Nov 5, 2010||Nov 19, 2013||International Business Machines Corporation||Allocation of energy budgets to individual partitions|
|US8595515||Jun 6, 2008||Nov 26, 2013||Google Inc.||Powering a data center|
|US8601287||Jun 6, 2008||Dec 3, 2013||Exaflop Llc||Computer and data center load determination|
|US8621248 *||May 4, 2011||Dec 31, 2013||Exaflop Llc||Load control in a data center|
|US8627128||Nov 9, 2010||Jan 7, 2014||International Business Machines Corporation||Power management for processing capacity upgrade on demand|
|US8645722||May 4, 2011||Feb 4, 2014||Exaflop Llc||Computer and data center load determination|
|US8694810||Sep 22, 2010||Apr 8, 2014||International Business Machines Corporation||Server power management with automatically-expiring server power allocations|
|US8700929 *||Jun 6, 2008||Apr 15, 2014||Exaflop Llc||Load control in a data center|
|US8707074||Aug 24, 2012||Apr 22, 2014||International Business Machines Corporation||Priority-based power capping in data processing systems|
|US8713334||Feb 4, 2013||Apr 29, 2014||Dell Products L.P.||Demand based power allocation|
|US8892728||Jul 21, 2011||Nov 18, 2014||Hewlett-Packard Development Company, L.P.||Automatic zone-based management of a data center|
|US8914157||Jun 20, 2013||Dec 16, 2014||Intel Corporation||Method and apparatus for dynamically allocating power in a data center|
|US8949646||Jun 6, 2008||Feb 3, 2015||Google Inc.||Data center load monitoring for utilizing an access power amount based on a projected peak power usage and a monitored power usage|
|US9009500||Jan 18, 2012||Apr 14, 2015||Google Inc.||Method of correlating power in a data center by fitting a function to a plurality of pairs of actual power draw values and estimated power draw values determined from monitored CPU utilization of a statistical sample of computers in the data center|
|US9026818||Aug 24, 2012||May 5, 2015||Lenovo Enterprise Solutions (Singapore) Pte. Ltd.||Priority-based power capping in data processing systems|
|US9146608||Dec 12, 2013||Sep 29, 2015||International Business Machines Corporation||Power management for processing capacity upgrade on demand|
|US9250684||Feb 25, 2015||Feb 2, 2016||Quanta Computer Inc.||Dynamic power capping of a subset of servers when a power consumption threshold is reached and allotting an amount of discretionary power to the servers that have power capping enabled|
|US9287710||May 8, 2012||Mar 15, 2016||Google Inc.||Supplying grid ancillary services using controllable loads|
|US9383791||Mar 9, 2015||Jul 5, 2016||Google Inc.||Accurate power allotment|
|US20090113221 *||Oct 29, 2007||Apr 30, 2009||Microsoft Corporation||Collaborative power sharing between computing devices|
|US20090319808 *||Jun 20, 2008||Dec 24, 2009||Dell Products, Lp||System and Method for Smoothing Power Reclamation of Blade Servers|
|US20100037070 *||Feb 11, 2010||Dell Products L.P.||Demand based power allocation|
|US20100037077 *||Feb 11, 2010||Vivek Kashyap||Multiple-node system power utilization management|
|US20100162013 *||Jul 22, 2009||Jun 24, 2010||Electronics And Telecommunications Research Institute||Remote power management system and method for cluster system|
|US20100332872 *||Jun 30, 2009||Dec 30, 2010||International Business Machines Corporation||Priority-Based Power Capping in Data Processing Systems|
|US20110144818 *||Jun 16, 2011||Cong Li||Method and apparatus for dynamically allocating power in a data center|
|US20130297089 *||Sep 12, 2012||Nov 7, 2013||Sheau-Wei J. Fu||Power management control system|
|U.S. Classification||713/300, 713/320, 700/295|
|International Classification||G05D3/12, G06F1/00, G06F1/32|
|Jun 27, 2006||AS||Assignment|
Owner name: HEWLETT-PACKARD DEVELOPMENT COMPANY, L.P., TEXAS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:GOODRUM, ALAN L.;TIPLEY, ROGER E.;REEL/FRAME:018054/0555;SIGNING DATES FROM 20060614 TO 20060621
Owner name: HEWLETT-PACKARD DEVELOPMENT COMPANY, L.P.,TEXAS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:GOODRUM, ALAN L.;TIPLEY, ROGER E.;SIGNING DATES FROM 20060614 TO 20060621;REEL/FRAME:018054/0555
Owner name: HEWLETT-PACKARD DEVELOPMENT COMPANY, L.P., TEXAS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:GOODRUM, ALAN L.;TIPLEY, ROGER E.;SIGNING DATES FROM 20060614 TO 20060621;REEL/FRAME:018054/0555
|Sep 24, 2013||FPAY||Fee payment|
Year of fee payment: 4
|Nov 9, 2015||AS||Assignment|
Owner name: HEWLETT PACKARD ENTERPRISE DEVELOPMENT LP, TEXAS
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:HEWLETT-PACKARD DEVELOPMENT COMPANY, L.P.;REEL/FRAME:037079/0001
Effective date: 20151027